Sparsity Based Reconstruction/Compressive Sensing in Medical Imaging
Michael (Miki) Lustig, Guanghong Chen, and Joshua D Trzasko
Abstract
Medical imaging has revolutionized diagnostic medicine. For over a century we can look inside the body without having to physically cut through. Clinicians today have an arsenal of different modalities that provide important information on both structure and function with new modalities and application constantly being developed.
Most medical imaging systems work by depositing energy into the body using an energy source. That energy encodes spatial positions and information of the tissue. Then, energy is received using detectors and its information content is decoded to produce images. For example, in Computed Tomography (CT) X-ray radiation is projected from various angles to produce an image of the spatial distribution of X-ray attenuation coefficients in the body, In Ultrasound, reflection of high frequency sound waves are used to form images of tissue interfaces. In MRI radio frequency pulses and magnetic fields are used to encode the spatial position of proton spins in water and create images of various properties of these protons in different tissues.
Image encoding and reconstruction are ubiquitous across all medical imaging modalities. With constant improvements in technology, imaging bottlenecks are shifting from hardware to limits in the ability to encode images. For example in MRI, rapid collection of data is limited by physical and physiological constraints on switching magnetic fields. In CT, it is constrained by radiation dose limits. Traditionally, encoding and and reconstruction did not assume anything about the content of an image (except perhaps field of view). However medical images, much like natural images taken with digital cameras are redundant. Encoding and reconstruction that incorporates image models can potentially provide much better imaging tradeoffs that result in higher quality, better signal to noise, higher resolution, less radiation dose, faster acquisition time, etc. An increasingly popular medical imaging model is sparse representation. Sparsity is a powerful model. On one hand it constrains the image to be simple, i.e., the image can be represented using only a few non-zero coefficients from a known dictionary. On the other hand, it also provides considerable degrees of freedom to represent almost any naturally occurring image features. Motivated by the success of lossy image compression using sparse representations and the recent introduction of compressed sensing, sparsity based reconstructions are rapidly finding their way into the clinic.
In this tutorial we will focus on two modalities: MRI in which sparsity based reconstructions are used to significantly accelerate scan time and provide images at unprecedented spatial and temporal resolution, and CT in which sparsity based reconstructions can lead to significant radiation dose reduction and superior diagnostic imagery.
Outline of the tutorial
- Principles of Magnetic Resonance Imaging
- Principles of Computed Tomography
- Sparse Models in Medical Imaging
- The application of Compressed Sensing in MRI and CT for fast imaging and radiation dose reduction
- Specific examples of clinical application of sparsity based reconstructions
- Current trends in sparsity based reconstruction in medical imaging
- Pitfalls in using sparsity baed reconstruction
- Future directions
Speakers
Michael (Miki) Lustig
Michael (Miki) Lustig is an Assistant Professor in EECS. He joined the faculty in Spring 2010. He received his Bsc in Electrical Engineering from the Technion, Israel Institute of Technology in 2002. He received his Msc and PhD in Electrical Engineering from Stanford University in 2004 and 2008 respectively. His research focuses on medical imaging, in particular Magnetic Resonance Imaging (MRI). More specifically, the application of compressed sensing to rapid and high-resolution MRI, MRI pulse sequence design, medical image reconstruction, inverse problems in medical imaging and sparse signal representation.
Guanghong Chen
Dr. Chen received his BS degree in physics in 1991, a MS degree in mathematics in 1997, and a PhD in semiconductor physics in 1999. His research has crossed many sub--disciplines inmathematics, condensed matter physics, optics, and string theory before he started his career as a medical physicist in 2002. He is currently a professor of Medical Physics and Radiology and the Director of X--ray CT Imaging Center at the University of Wisconsin--Madison. His current research activities focus on novel image reconstruction algorithm development and new x--ray CT image acquisition system development. Using the developed new algorithms and new data acquisition devices, Dr. Chen and his basic science and clinical research team are working on cutting--edge clinical applications of x--ray CT in cancer screening, diagnosis, minimally invasive cerebral and cardiac interventions, and image--guided radiation therapy. Dr. Chen has published more than 110 papers, delivered more than 60 invited talks at major national and international conferences, and had more than 20 US and international patents.
Joshua D Trzasko
Trzasko (Sí99) received the B.S. degree in electrical and computer engineering from Cornell University, Ithaca, NY, in 2003, and the Ph.D. degree in biomedical engineering from Mayo Graduate School, Rochester, MN in 2009. In the summer of 2002, he interned at General Electricís Global Research Center, Niskayuna, NY, in the MRI division. Dr. Trzasko is currently a research fellow in the Department of Physiology and Biomedical Engineering at Mayo Clinic. He is interested in statistical signal processing applications in medical imaging, including estimation and correction problems in MRI and low-dose CT imaging.